@oh-my-pi/pi-coding-agent 15.7.1 → 15.7.2

package/CHANGELOG.md

@@ -2,6 +2,26 @@
 
 ## [Unreleased]
 
+## [15.7.2] - 2026-05-31
+### Added
+
+- Added `providers.autoThinkingModel` setting so users can choose the `auto` thinking classifier backend (online smol or local tiny-memory model)
+- Added an `auto` thinking level that classifies each real user turn and resolves to a concrete low-through-xhigh effort, with online smol classification by default and an opt-in local on-device classifier.
+
+### Changed
+
+- Updated the interactive thinking selectors in model/model-role pickers and ACP thinking options to include `auto` as a selectable level
+- Updated footer and status-line rendering to show `auto` while auto-thinking is being resolved and `auto → <level>` once it resolves
+- Changed the local tiny-model device default to CPU on every platform; explicit `providers.tinyModelDevice` / `PI_TINY_DEVICE` values still opt into accelerated ONNX providers.
+
+### Fixed
+
+- Prevented auto-thinking classification from running on non-user synthetic turns and non-reasoning models, keeping the session on its provisional concrete effort
+- Added a bounded auto-thinking classification path that falls back to the provisional effort on failures/timeouts so prompts continue without interruption
+- Bypassed auto classifier for `ultrathink` prompts and resolved directly to the highest supported auto effort
+- Fixed the JavaScript `eval` kernel crashing the whole process with a segfault (`SIGTRAP`, `getImportedModule` on a null record) when imported code reached a local module whose relative-import graph contains a cycle — e.g. `await import("…/edit/streaming.ts")`, or any workspace path with cyclic re-exports. The `LocalModuleLoader` linked and evaluated each local module individually inside the recursive `vm.SourceTextModule` linker callback, which re-entered Bun's `node:vm` module linker mid-instantiation and detonated JSC on the first cycle. The loader now constructs the entire local module graph first and drives a single `link()` + `evaluate()` from the graph root, so cyclic graphs instantiate in one pass; external (`node_modules`) modules stay eagerly loaded since they carry no imports and cannot form a cycle.
+- Fixed the streaming `edit` preview rendering a blank box for hashline edits whose payload sits on the trailing in-flight line (the common single-op `replace`/`insert` case). The preview path trimmed that still-typing line before diffing, so a single-payload op collapsed to a "No changes" result — shown as an empty box — for almost the entire stream. Hashline previews now feed the raw in-flight text through `applyPartialTo`, whose streaming-tolerant parser drops a payload-less trailing op and projects a partially-typed payload line as it grows, so the diff appears and fills in live. Transient errors from the actively-typed trailing section are also suppressed while streaming (regardless of section count) so a mid-typed op can't wipe an already-good preview frame; real errors still surface once args are complete.
+
 ## [15.7.0] - 2026-05-31
 
 ### Added

package/dist/types/auto-thinking/classifier.d.ts

@@ -0,0 +1,35 @@
+/**
+ * Per-prompt difficulty classifier for the `auto` thinking level.
+ *
+ * Picks a coding-difficulty bucket for a user prompt and maps it to a concrete
+ * {@link Effort}, clamped into the active model's supported range (never below
+ * {@link Effort.Low}). Two backends, selected by `providers.autoThinkingModel`:
+ *
+ * - `online` (default): a smol model classifies into `low|medium|high|xhigh`.
+ * - a local key: an on-device memory model classifies into the coarser
+ *   `trivial|moderate|hard` scheme (3-class is more reliable than 4-way ordinal
+ *   on sub-2B models), mapped to `low|high|xhigh`.
+ *
+ * Throws on any failure (no model, no key, unparseable output, abort/timeout);
+ * the caller falls back to a concrete level and continues the turn.
+ */
+import { Effort, type Model } from "@oh-my-pi/pi-ai";
+import type { ModelRegistry } from "../config/model-registry";
+import type { Settings } from "../config/settings";
+export interface ClassifyDifficultyDeps {
+    settings: Settings;
+    registry: ModelRegistry;
+    model: Model;
+    sessionId?: string;
+    signal?: AbortSignal;
+    metadataResolver?: (provider: string) => Record<string, unknown> | undefined;
+}
+/**
+ * Classify `promptText` and return a concrete effort clamped to `deps.model`.
+ * @throws when the backend cannot produce a usable classification.
+ */
+export declare function classifyDifficulty(promptText: string, deps: ClassifyDifficultyDeps): Promise<Effort>;
+/** Map the online 4-way level keyword to an {@link Effort}; earliest match wins. */
+export declare function parseDifficultyLevel(text: string): Effort | undefined;
+/** Map the local 3-way bucket keyword to an {@link Effort}; earliest match wins. */
+export declare function parseDifficultyBucket(text: string): Effort | undefined;

package/dist/types/config/settings-schema.d.ts

@@ -704,13 +704,13 @@ export declare const SETTINGS_SCHEMA: {
     };
     readonly defaultThinkingLevel: {
         readonly type: "enum";
-        readonly values: readonly import("@oh-my-pi/pi-ai").Effort[];
+        readonly values: readonly [...import("@oh-my-pi/pi-ai").Effort[], "auto"];
         readonly default: "high";
         readonly ui: {
             readonly tab: "model";
             readonly label: "Thinking Level";
             readonly description: "Reasoning depth for thinking-capable models";
-            readonly options: readonly import("../thinking").ThinkingLevelMetadata[];
+            readonly options: readonly [import("../thinking").ConfiguredThinkingLevelMetadata, ...import("../thinking").ThinkingLevelMetadata[]];
         };
     };
     readonly hideThinkingBlock: {
@@ -3385,11 +3385,11 @@ export declare const SETTINGS_SCHEMA: {
         readonly ui: {
             readonly tab: "providers";
             readonly label: "Tiny Model Device";
-            readonly description: "ONNX execution provider for local tiny models (titles + memory). Default picks DirectML on Windows, CUDA on Linux x64, CPU elsewhere. The PI_TINY_DEVICE env var overrides this.";
+            readonly description: "ONNX execution provider for local tiny models (titles + memory). Default uses CPU-only inference. The PI_TINY_DEVICE env var overrides this.";
             readonly options: readonly [{
                 readonly value: "default";
                 readonly label: "Default";
-                readonly description: "DirectML on Windows, CUDA on Linux x64, CPU elsewhere";
+                readonly description: "CPU-only inference";
             }, {
                 readonly value: "gpu";
                 readonly label: "GPU";
@@ -3532,6 +3532,26 @@ export declare const SETTINGS_SCHEMA: {
             })[];
         };
     };
+    readonly "providers.autoThinkingModel": {
+        readonly type: "enum";
+        readonly values: readonly ["online", "qwen3-1.7b", "gemma-3-1b", "qwen2.5-1.5b", "lfm2-1.2b"];
+        readonly default: "online";
+        readonly ui: {
+            readonly tab: "model";
+            readonly label: "Auto Thinking Model";
+            readonly description: "Difficulty classifier for the `auto` thinking level: online smol by default, or a local on-device model";
+            readonly condition: "autoThinkingActive";
+            readonly options: ({
+                value: "online";
+                label: string;
+                description: string;
+            } | {
+                value: "gemma-3-1b" | "lfm2-1.2b" | "qwen2.5-1.5b" | "qwen3-1.7b";
+                label: "Gemma 3 1B" | "LFM2 1.2B" | "Qwen2.5 1.5B" | "Qwen3 1.7B";
+                description: "Best consolidation/dedup; lighter footprint, but leaks small talk during extraction." | "Best extraction granularity (atomic facts); weaker consolidation." | "Fastest load; solid all-rounder, slightly noisier extraction labels." | "Recommended; most disciplined extraction (ignores chit-chat), good consolidation, about 1.1 GB cached.";
+            })[];
+        };
+    };
     readonly "providers.kimiApiFormat": {
         readonly type: "enum";
         readonly values: readonly ["openai", "anthropic"];

package/dist/types/edit/hashline/diff.d.ts

@@ -17,6 +17,12 @@ export interface HashlineDiffOptions {
      * preview path only.
      */
     streaming?: boolean;
+    /**
+     * Skip snapshot-tag validation. Streaming previews use this so transient
+     * stale/missing tags do not flash re-read errors while the model is still
+     * authoring input; the final apply path still validates through Patcher.
+     */
+    skipHashValidation?: boolean;
 }
 export declare function computeHashlineSectionDiff(section: PatchSection, cwd: string, snapshots: SnapshotStore, options?: HashlineDiffOptions): Promise<{
     diff: string;

package/dist/types/modes/components/model-selector.d.ts

@@ -1,12 +1,13 @@
-import { ThinkingLevel } from "@oh-my-pi/pi-agent-core";
 import { type Model } from "@oh-my-pi/pi-ai";
 import { Container, Input, type TUI } from "@oh-my-pi/pi-tui";
 import type { ModelRegistry } from "../../config/model-registry";
 import type { Settings } from "../../config/settings";
+import { type ConfiguredThinkingLevel } from "../../thinking";
 interface ScopedModelItem {
     model: Model;
     thinkingLevel?: string;
 }
+type RoleSelectCallback = (model: Model, role: string | null, thinkingLevel?: ConfiguredThinkingLevel, selector?: string) => void;
 /**
  * Component that renders a model selector with provider tabs and context menu.
  * - Tab/Arrow Left/Right: Switch between provider tabs
@@ -16,7 +17,7 @@ interface ScopedModelItem {
  */
 export declare class ModelSelectorComponent extends Container {
     #private;
-    constructor(tui: TUI, _currentModel: Model | undefined, settings: Settings, modelRegistry: ModelRegistry, scopedModels: ReadonlyArray<ScopedModelItem>, onSelect: (model: Model, role: string | null, thinkingLevel?: ThinkingLevel, selector?: string) => void, onCancel: () => void, options?: {
+    constructor(tui: TUI, _currentModel: Model | undefined, settings: Settings, modelRegistry: ModelRegistry, scopedModels: ReadonlyArray<ScopedModelItem>, onSelect: RoleSelectCallback, onCancel: () => void, options?: {
         temporaryOnly?: boolean;
         initialSearchInput?: string;
     });

package/dist/types/modes/theme/theme.d.ts

@@ -6,7 +6,7 @@ export type SymbolPreset = "unicode" | "nerd" | "ascii";
 /**
  * All available symbol keys organized by category.
  */
-export type SymbolKey = "status.success" | "status.error" | "status.warning" | "status.info" | "status.pending" | "status.disabled" | "status.enabled" | "status.running" | "status.shadowed" | "status.aborted" | "nav.cursor" | "nav.selected" | "nav.expand" | "nav.collapse" | "nav.back" | "tree.branch" | "tree.last" | "tree.vertical" | "tree.horizontal" | "tree.hook" | "boxRound.topLeft" | "boxRound.topRight" | "boxRound.bottomLeft" | "boxRound.bottomRight" | "boxRound.horizontal" | "boxRound.vertical" | "boxSharp.topLeft" | "boxSharp.topRight" | "boxSharp.bottomLeft" | "boxSharp.bottomRight" | "boxSharp.horizontal" | "boxSharp.vertical" | "boxSharp.cross" | "boxSharp.teeDown" | "boxSharp.teeUp" | "boxSharp.teeRight" | "boxSharp.teeLeft" | "sep.powerline" | "sep.powerlineThin" | "sep.powerlineLeft" | "sep.powerlineRight" | "sep.powerlineThinLeft" | "sep.powerlineThinRight" | "sep.block" | "sep.space" | "sep.asciiLeft" | "sep.asciiRight" | "sep.dot" | "sep.slash" | "sep.pipe" | "icon.model" | "icon.plan" | "icon.goal" | "icon.pause" | "icon.loop" | "icon.folder" | "icon.scratchFolder" | "icon.file" | "icon.git" | "icon.branch" | "icon.pr" | "icon.tokens" | "icon.context" | "icon.cost" | "icon.time" | "icon.pi" | "icon.agents" | "icon.cache" | "icon.input" | "icon.output" | "icon.host" | "icon.session" | "icon.package" | "icon.warning" | "icon.rewind" | "icon.auto" | "icon.fast" | "icon.extensionSkill" | "icon.extensionTool" | "icon.extensionSlashCommand" | "icon.extensionMcp" | "icon.extensionRule" | "icon.extensionHook" | "icon.extensionPrompt" | "icon.extensionContextFile" | "icon.extensionInstruction" | "icon.mic" | "thinking.minimal" | "thinking.low" | "thinking.medium" | "thinking.high" | "thinking.xhigh" | "checkbox.checked" | "checkbox.unchecked" | "format.bullet" | "format.dash" | "format.bracketLeft" | "format.bracketRight" | "md.quoteBorder" | "md.hrChar" | "md.bullet" | "md.colorSwatch" | "lang.default" | "lang.typescript" | "lang.javascript" | "lang.python" | "lang.rust" | "lang.go" | "lang.java" | "lang.c" | "lang.cpp" | "lang.csharp" | "lang.ruby" | "lang.php" | "lang.swift" | "lang.kotlin" | "lang.shell" | "lang.html" | "lang.css" | "lang.json" | "lang.yaml" | "lang.markdown" | "lang.sql" | "lang.docker" | "lang.lua" | "lang.text" | "lang.env" | "lang.toml" | "lang.xml" | "lang.ini" | "lang.conf" | "lang.log" | "lang.csv" | "lang.tsv" | "lang.image" | "lang.pdf" | "lang.archive" | "lang.binary" | "tab.appearance" | "tab.model" | "tab.interaction" | "tab.context" | "tab.editing" | "tab.tools" | "tab.memory" | "tab.tasks" | "tab.providers";
+export type SymbolKey = "status.success" | "status.error" | "status.warning" | "status.info" | "status.pending" | "status.disabled" | "status.enabled" | "status.running" | "status.shadowed" | "status.aborted" | "nav.cursor" | "nav.selected" | "nav.expand" | "nav.collapse" | "nav.back" | "tree.branch" | "tree.last" | "tree.vertical" | "tree.horizontal" | "tree.hook" | "boxRound.topLeft" | "boxRound.topRight" | "boxRound.bottomLeft" | "boxRound.bottomRight" | "boxRound.horizontal" | "boxRound.vertical" | "boxSharp.topLeft" | "boxSharp.topRight" | "boxSharp.bottomLeft" | "boxSharp.bottomRight" | "boxSharp.horizontal" | "boxSharp.vertical" | "boxSharp.cross" | "boxSharp.teeDown" | "boxSharp.teeUp" | "boxSharp.teeRight" | "boxSharp.teeLeft" | "sep.powerline" | "sep.powerlineThin" | "sep.powerlineLeft" | "sep.powerlineRight" | "sep.powerlineThinLeft" | "sep.powerlineThinRight" | "sep.block" | "sep.space" | "sep.asciiLeft" | "sep.asciiRight" | "sep.dot" | "sep.slash" | "sep.pipe" | "icon.model" | "icon.plan" | "icon.goal" | "icon.pause" | "icon.loop" | "icon.folder" | "icon.scratchFolder" | "icon.file" | "icon.git" | "icon.branch" | "icon.pr" | "icon.tokens" | "icon.context" | "icon.cost" | "icon.time" | "icon.pi" | "icon.agents" | "icon.cache" | "icon.input" | "icon.output" | "icon.host" | "icon.session" | "icon.package" | "icon.warning" | "icon.rewind" | "icon.auto" | "icon.fast" | "icon.extensionSkill" | "icon.extensionTool" | "icon.extensionSlashCommand" | "icon.extensionMcp" | "icon.extensionRule" | "icon.extensionHook" | "icon.extensionPrompt" | "icon.extensionContextFile" | "icon.extensionInstruction" | "icon.mic" | "thinking.minimal" | "thinking.low" | "thinking.medium" | "thinking.high" | "thinking.xhigh" | "thinking.autoPending" | "checkbox.checked" | "checkbox.unchecked" | "format.bullet" | "format.dash" | "format.bracketLeft" | "format.bracketRight" | "md.quoteBorder" | "md.hrChar" | "md.bullet" | "md.colorSwatch" | "lang.default" | "lang.typescript" | "lang.javascript" | "lang.python" | "lang.rust" | "lang.go" | "lang.java" | "lang.c" | "lang.cpp" | "lang.csharp" | "lang.ruby" | "lang.php" | "lang.swift" | "lang.kotlin" | "lang.shell" | "lang.html" | "lang.css" | "lang.json" | "lang.yaml" | "lang.markdown" | "lang.sql" | "lang.docker" | "lang.lua" | "lang.text" | "lang.env" | "lang.toml" | "lang.xml" | "lang.ini" | "lang.conf" | "lang.log" | "lang.csv" | "lang.tsv" | "lang.image" | "lang.pdf" | "lang.archive" | "lang.binary" | "tab.appearance" | "tab.model" | "tab.interaction" | "tab.context" | "tab.editing" | "tab.tools" | "tab.memory" | "tab.tasks" | "tab.providers";
 export type SpinnerType = "status" | "activity";
 export type ThemeColor = "accent" | "border" | "borderAccent" | "borderMuted" | "success" | "error" | "warning" | "muted" | "dim" | "text" | "thinkingText" | "userMessageText" | "customMessageText" | "customMessageLabel" | "toolTitle" | "toolOutput" | "mdHeading" | "mdLink" | "mdLinkUrl" | "mdCode" | "mdCodeBlock" | "mdCodeBlockBorder" | "mdQuote" | "mdQuoteBorder" | "mdHr" | "mdListBullet" | "toolDiffAdded" | "toolDiffRemoved" | "toolDiffContext" | "syntaxComment" | "syntaxKeyword" | "syntaxFunction" | "syntaxVariable" | "syntaxString" | "syntaxNumber" | "syntaxType" | "syntaxOperator" | "syntaxPunctuation" | "thinkingOff" | "thinkingMinimal" | "thinkingLow" | "thinkingMedium" | "thinkingHigh" | "thinkingXhigh" | "bashMode" | "pythonMode" | "statusLineSep" | "statusLineModel" | "statusLinePath" | "statusLineGitClean" | "statusLineGitDirty" | "statusLineContext" | "statusLineSpend" | "statusLineStaged" | "statusLineDirty" | "statusLineUntracked" | "statusLineOutput" | "statusLineCost" | "statusLineSubagents";
 /** Check if a string is a valid ThemeColor value */
@@ -161,6 +161,7 @@ export declare class Theme {
         medium: string;
         high: string;
         xhigh: string;
+        autoPending: string;
     };
     get checkbox(): {
         checked: string;

package/dist/types/sdk.d.ts

@@ -19,6 +19,7 @@ import { AgentSession } from "./session/agent-session";
 import { AuthStorage } from "./session/auth-storage";
 import { SessionManager } from "./session/session-manager";
 import { type BuildSystemPromptResult } from "./system-prompt";
+import { type ConfiguredThinkingLevel } from "./thinking";
 import { BashTool, BUILTIN_TOOLS, createTools, EditTool, EvalTool, FindTool, HIDDEN_TOOLS, type LspStartupServerInfo, loadSshTool, ReadTool, ResolveTool, SearchTool, type Tool, type ToolSession, WebSearchTool, WriteTool } from "./tools";
 import { EventBus } from "./utils/event-bus";
 import { type WorkspaceTree } from "./workspace-tree";
@@ -39,7 +40,7 @@ export interface CreateAgentSessionOptions {
      * Used when model lookup is deferred because extension-provided models aren't registered yet. */
     modelPattern?: string;
     /** Thinking selector. Default: from settings, else unset */
-    thinkingLevel?: ThinkingLevel;
+    thinkingLevel?: ConfiguredThinkingLevel;
     /** Models available for cycling (Ctrl+P in interactive mode) */
     scopedModels?: Array<{
         model: Model;

package/dist/types/session/agent-session.d.ts

@@ -15,7 +15,8 @@
 import type { InMemorySnapshotStore } from "@oh-my-pi/hashline";
 import { type Agent, type AgentEvent, type AgentMessage, type AgentState, type AgentTool, ThinkingLevel } from "@oh-my-pi/pi-agent-core";
 import { type CompactionResult } from "@oh-my-pi/pi-agent-core/compaction";
-import type { AssistantMessage, Effort, ImageContent, Message, MessageAttribution, Model, ProviderSessionState, ServiceTier, SimpleStreamOptions, TextContent, ToolChoice, UsageReport } from "@oh-my-pi/pi-ai";
+import type { AssistantMessage, ImageContent, Message, MessageAttribution, Model, ProviderSessionState, ServiceTier, SimpleStreamOptions, TextContent, ToolChoice, UsageReport } from "@oh-my-pi/pi-ai";
+import { Effort } from "@oh-my-pi/pi-ai";
 import { type AsyncJob, type AsyncJobDeliveryState, AsyncJobManager } from "../async";
 import type { Rule } from "../capability/rule";
 import { type ModelRegistry } from "../config/model-registry";
@@ -39,6 +40,7 @@ import { type MnemosyneSessionState } from "../mnemosyne/state";
 import type { PlanModeState } from "../plan-mode/state";
 import { type AgentRegistry } from "../registry/agent-registry";
 import { type SecretObfuscator } from "../secrets/obfuscator";
+import { type ConfiguredThinkingLevel } from "../thinking";
 import { type DiscoverableTool, type DiscoverableToolSearchIndex } from "../tool-discovery/tool-index";
 import type { CheckpointState } from "../tools/checkpoint";
 import { type TodoItem, type TodoPhase } from "../tools/todo-write";
@@ -102,6 +104,10 @@ export type AgentSessionEvent = AgentEvent | {
 } | {
     type: "thinking_level_changed";
     thinkingLevel: ThinkingLevel | undefined;
+    /** The user-configured selector when it differs from the effective level (e.g. `auto`). */
+    configured?: ConfiguredThinkingLevel;
+    /** The level `auto` resolved to this turn, once classified. */
+    resolved?: Effort;
 } | {
     type: "goal_updated";
     goal: Goal | null;
@@ -125,7 +131,7 @@ export interface AgentSessionConfig {
         thinkingLevel?: ThinkingLevel;
     }>;
     /** Initial session thinking selector. */
-    thinkingLevel?: ThinkingLevel;
+    thinkingLevel?: ConfiguredThinkingLevel;
     /** Prompt templates for expansion */
     promptTemplates?: PromptTemplate[];
     /** File-based slash commands for expansion */
@@ -364,8 +370,14 @@ export declare class AgentSession {
     get state(): AgentState;
     /** Current model (may be undefined if not yet selected) */
     get model(): Model | undefined;
-    /** Current thinking level */
+    /** Effective thinking level applied to the agent (the resolved level when `auto`). */
     get thinkingLevel(): ThinkingLevel | undefined;
+    /** The selector the user configured: `auto` when auto mode is active, else the effective level. */
+    configuredThinkingLevel(): ConfiguredThinkingLevel | undefined;
+    /** True when `auto` thinking mode is active. */
+    get isAutoThinking(): boolean;
+    /** The level `auto` resolved to for the current turn (undefined until classified). */
+    autoResolvedThinkingLevel(): Effort | undefined;
     get serviceTier(): ServiceTier | undefined;
     /** Whether agent is currently streaming a response */
     get isStreaming(): boolean;
@@ -654,15 +666,16 @@ export declare class AgentSession {
      */
     getAvailableModels(): Model[];
     /**
-     * Set thinking level.
-     * Saves the effective metadata-clamped level to session and settings only if it changes.
+     * Set the thinking level. `auto` enables per-turn classification (session-level,
+     * never written to the session log); a concrete level clears auto. The effective
+     * metadata-clamped level is saved to the session/settings only when it changes.
      */
-    setThinkingLevel(level: ThinkingLevel | undefined, persist?: boolean): void;
+    setThinkingLevel(level: ConfiguredThinkingLevel | undefined, persist?: boolean): void;
     /**
-     * Cycle to next thinking level.
-     * @returns New level, or undefined if model doesn't support thinking
+     * Cycle to next thinking level: off → auto → minimal..xhigh → off.
+     * @returns New selector, or undefined if model doesn't support thinking
      */
-    cycleThinkingLevel(): ThinkingLevel | undefined;
+    cycleThinkingLevel(): ConfiguredThinkingLevel | undefined;
     /**
      * True when *any* fast-mode-granting service tier is configured, regardless
      * of whether the active model's provider actually realizes it. Used by the

package/dist/types/thinking.d.ts

@@ -1,5 +1,5 @@
 import { type ResolvedThinkingLevel, ThinkingLevel } from "@oh-my-pi/pi-agent-core";
-import { type Effort, type Model } from "@oh-my-pi/pi-ai";
+import { Effort, type Model } from "@oh-my-pi/pi-ai";
 /**
  * Metadata used to render thinking selector values in the coding-agent UI.
  */
@@ -28,3 +28,41 @@ export declare function toReasoningEffort(level: ThinkingLevel | undefined): Eff
  * Resolves a selector against the current model while preserving explicit "off".
  */
 export declare function resolveThinkingLevelForModel(model: Model | undefined, level: ThinkingLevel | undefined): ResolvedThinkingLevel | undefined;
+/**
+ * Sentinel selector for the coding-agent "auto" thinking mode. Kept entirely
+ * inside the coding-agent layer: it is never an {@link Effort} or
+ * {@link ThinkingLevel}, so provider mapping/clamping keeps seeing concrete
+ * efforts. The session resolves `auto` to a concrete effort each turn.
+ */
+export declare const AUTO_THINKING: "auto";
+/** A thinking selector as configured by the user — a concrete level or `auto`. */
+export type ConfiguredThinkingLevel = ThinkingLevel | typeof AUTO_THINKING;
+/** Metadata used to render the `auto` selector value alongside concrete levels. */
+export interface ConfiguredThinkingLevelMetadata {
+    value: ConfiguredThinkingLevel;
+    label: string;
+    description: string;
+}
+/**
+ * Parses a configured thinking selector, accepting `auto` in addition to every
+ * value {@link parseThinkingLevel} accepts. {@link parseThinkingLevel} itself
+ * stays strict so model-suffix parsing (`model:high`) keeps rejecting `auto`.
+ */
+export declare function parseConfiguredThinkingLevel(value: string | null | undefined): ConfiguredThinkingLevel | undefined;
+/** Returns display metadata for a configured selector, including `auto`. */
+export declare function getConfiguredThinkingLevelMetadata(level: ConfiguredThinkingLevel): ConfiguredThinkingLevelMetadata;
+/**
+ * Resolves an auto-classified effort against the active model's supported
+ * range. Unlike {@link clampThinkingLevelForModel}, `auto` never resolves below
+ * {@link Effort.Low}: the eligible pool is the model's supported efforts at or
+ * above Low (falling back to the full supported set only when the model maxes
+ * out below Low). Within that pool the request snaps to the highest level not
+ * exceeding it, or the pool minimum when the request is below the pool.
+ */
+export declare function clampAutoThinkingEffort(model: Model | undefined, effort: Effort): Effort;
+/**
+ * The provisional concrete level shown while `auto` is configured but before a
+ * turn has been classified. Prefers the model's `defaultLevel`, otherwise High,
+ * clamped into the auto range. Returns `undefined` for non-reasoning models.
+ */
+export declare function resolveProvisionalAutoLevel(model: Model | undefined): Effort | undefined;

package/dist/types/tiny/device.d.ts

@@ -7,7 +7,7 @@ export interface TinyModelDevicePreference {
 export declare function normalizeTinyModelDevice(value: string | undefined): TinyModelDevice | undefined;
 export declare function resolveTinyModelDevicePreference(value?: string | undefined): TinyModelDevicePreference;
 export declare function tinyModelDeviceLoadOrder(preference: TinyModelDevicePreference): readonly TinyModelDevice[];
-/** Sentinel `providers.tinyModelDevice` value meaning "use the built-in platform default". */
+/** Sentinel `providers.tinyModelDevice` value meaning "use the built-in CPU default". */
 export declare const TINY_MODEL_DEVICE_DEFAULT = "default";
 /** Accepted values for the `providers.tinyModelDevice` setting (validation + UI). */
 export declare const TINY_MODEL_DEVICE_SETTING_VALUES: readonly ["default", "gpu", "cpu", "metal", "webgpu", "cuda", "dml", "coreml", "auto", "wasm", "webnn", "webnn-gpu", "webnn-cpu", "webnn-npu"];
@@ -15,7 +15,7 @@ export declare const TINY_MODEL_DEVICE_SETTING_VALUES: readonly ["default", "gpu
 export declare const TINY_MODEL_DEVICE_SETTING_OPTIONS: readonly [{
     readonly value: "default";
     readonly label: "Default";
-    readonly description: "DirectML on Windows, CUDA on Linux x64, CPU elsewhere";
+    readonly description: "CPU-only inference";
 }, {
     readonly value: "gpu";
     readonly label: "GPU";
@@ -72,7 +72,7 @@ export declare const TINY_MODEL_DEVICE_SETTING_OPTIONS: readonly [{
 /**
  * Map a `providers.tinyModelDevice` setting value onto a `PI_TINY_DEVICE` env
  * value for the worker. Returns `undefined` for the default sentinel so the
- * worker keeps its built-in platform default; the worker still validates the
+ * worker keeps its built-in CPU default; the worker still validates the
  * forwarded value via {@link normalizeTinyModelDevice}.
  */
 export declare function tinyModelDeviceSettingToEnv(value: string | undefined): string | undefined;

package/dist/types/tiny/models.d.ts

@@ -183,3 +183,22 @@ export declare const TINY_LOCAL_MODELS: readonly [{
     readonly description: "Fastest load; solid all-rounder, slightly noisier extraction labels.";
     readonly contextNote: "Use when local startup cost is the priority.";
 }];
+/**
+ * Difficulty-classifier model for the `auto` thinking level. Defaults to the
+ * online smol path; the local options reuse the memory-model registry because
+ * the shared worker's `complete()` only accepts memory local keys, and the
+ * 1B+ memory models classify coding difficulty far more reliably than the
+ * sub-1B title models.
+ */
+export declare const ONLINE_AUTO_THINKING_MODEL_KEY = "online";
+export declare const AUTO_THINKING_MODEL_VALUES: readonly ["online", "qwen3-1.7b", "gemma-3-1b", "qwen2.5-1.5b", "lfm2-1.2b"];
+export type AutoThinkingModelKey = TinyMemoryModelKey;
+export declare const AUTO_THINKING_MODEL_OPTIONS: ({
+    value: "online";
+    label: string;
+    description: string;
+} | {
+    value: "gemma-3-1b" | "lfm2-1.2b" | "qwen2.5-1.5b" | "qwen3-1.7b";
+    label: "Gemma 3 1B" | "LFM2 1.2B" | "Qwen2.5 1.5B" | "Qwen3 1.7B";
+    description: "Best consolidation/dedup; lighter footprint, but leaks small talk during extraction." | "Best extraction granularity (atomic facts); weaker consolidation." | "Fastest load; solid all-rounder, slightly noisier extraction labels." | "Recommended; most disciplined extraction (ignores chit-chat), good consolidation, about 1.1 GB cached.";
+})[];

package/package.json

@@ -1,7 +1,7 @@
 {
 	"type": "module",
 	"name": "@oh-my-pi/pi-coding-agent",
-	"version": "15.7.1",
+	"version": "15.7.2",
 	"description": "Coding agent CLI with read, bash, edit, write tools and session management",
 	"homepage": "https://omp.sh",
 	"author": "Can Boluk",
@@ -47,14 +47,14 @@
 		"@agentclientprotocol/sdk": "0.22.1",
 		"@babel/parser": "^7.29.7",
 		"@mozilla/readability": "^0.6.0",
-		"@oh-my-pi/hashline": "15.7.1",
-		"@oh-my-pi/omp-stats": "15.7.1",
-		"@oh-my-pi/pi-agent-core": "15.7.1",
-		"@oh-my-pi/pi-ai": "15.7.1",
-		"@oh-my-pi/pi-mnemosyne": "15.7.1",
-		"@oh-my-pi/pi-natives": "15.7.1",
-		"@oh-my-pi/pi-tui": "15.7.1",
-		"@oh-my-pi/pi-utils": "15.7.1",
+		"@oh-my-pi/hashline": "15.7.2",
+		"@oh-my-pi/omp-stats": "15.7.2",
+		"@oh-my-pi/pi-agent-core": "15.7.2",
+		"@oh-my-pi/pi-ai": "15.7.2",
+		"@oh-my-pi/pi-mnemosyne": "15.7.2",
+		"@oh-my-pi/pi-natives": "15.7.2",
+		"@oh-my-pi/pi-tui": "15.7.2",
+		"@oh-my-pi/pi-utils": "15.7.2",
 		"@puppeteer/browsers": "^3.0.4",
 		"@types/turndown": "5.0.6",
 		"@xterm/headless": "^6.0.0",

package/src/auto-thinking/classifier.ts

@@ -0,0 +1,180 @@
+/**
+ * Per-prompt difficulty classifier for the `auto` thinking level.
+ *
+ * Picks a coding-difficulty bucket for a user prompt and maps it to a concrete
+ * {@link Effort}, clamped into the active model's supported range (never below
+ * {@link Effort.Low}). Two backends, selected by `providers.autoThinkingModel`:
+ *
+ * - `online` (default): a smol model classifies into `low|medium|high|xhigh`.
+ * - a local key: an on-device memory model classifies into the coarser
+ *   `trivial|moderate|hard` scheme (3-class is more reliable than 4-way ordinal
+ *   on sub-2B models), mapped to `low|high|xhigh`.
+ *
+ * Throws on any failure (no model, no key, unparseable output, abort/timeout);
+ * the caller falls back to a concrete level and continues the turn.
+ */
+import { type AssistantMessage, completeSimple, Effort, type Model } from "@oh-my-pi/pi-ai";
+import { prompt } from "@oh-my-pi/pi-utils";
+import type { ModelRegistry } from "../config/model-registry";
+import { resolveRoleSelection } from "../config/model-resolver";
+import type { Settings } from "../config/settings";
+import difficultySystemPrompt from "../prompts/system/auto-thinking-difficulty.md" with { type: "text" };
+import difficultyLocalPrompt from "../prompts/system/auto-thinking-difficulty-local.md" with { type: "text" };
+import { clampAutoThinkingEffort } from "../thinking";
+import { isTinyMemoryLocalModelKey, ONLINE_AUTO_THINKING_MODEL_KEY } from "../tiny/models";
+import { tinyModelClient } from "../tiny/title-client";
+
+const DIFFICULTY_SYSTEM_PROMPT = prompt.render(difficultySystemPrompt);
+
+/** Upper bound on prompt characters fed to the classifier. */
+const MAX_INPUT_CHARS = 6000;
+const HEAD_CHARS = 4000;
+const TAIL_CHARS = 2000;
+/** The answer is a single word; keep budgets tiny for non-reasoning backends. */
+const ANSWER_MAX_TOKENS = 8;
+/**
+ * Reasoning backends ignore `disableReasoning` on some providers, so reserve
+ * enough output room for the keyword to still land after unavoidable thinking.
+ */
+const REASONING_SAFE_MAX_TOKENS = 1024;
+
+export interface ClassifyDifficultyDeps {
+	settings: Settings;
+	registry: ModelRegistry;
+	model: Model;
+	sessionId?: string;
+	signal?: AbortSignal;
+	metadataResolver?: (provider: string) => Record<string, unknown> | undefined;
+}
+
+/**
+ * Classify `promptText` and return a concrete effort clamped to `deps.model`.
+ * @throws when the backend cannot produce a usable classification.
+ */
+export async function classifyDifficulty(promptText: string, deps: ClassifyDifficultyDeps): Promise<Effort> {
+	const backend = deps.settings.get("providers.autoThinkingModel");
+	const input = prepareClassifierInput(promptText);
+	const effort =
+		backend === ONLINE_AUTO_THINKING_MODEL_KEY
+			? await classifyOnline(input, deps)
+			: await classifyLocal(input, backend, deps);
+	return clampAutoThinkingEffort(deps.model, effort);
+}
+
+async function classifyOnline(input: string, deps: ClassifyDifficultyDeps): Promise<Effort> {
+	const resolved = resolveRoleSelection(["smol"], deps.settings, deps.registry.getAvailable(), deps.registry);
+	const model = resolved?.model;
+	if (!model) {
+		throw new Error("auto-thinking: no smol model available for classification");
+	}
+	const apiKey = await deps.registry.getApiKey(model, deps.sessionId);
+	if (!apiKey) {
+		throw new Error(`auto-thinking: no API key for ${model.provider}/${model.id}`);
+	}
+	// Resolve metadata after getApiKey so the session-sticky credential is recorded first.
+	const metadata = deps.metadataResolver?.(model.provider);
+	const maxTokens = model.reasoning ? Math.max(ANSWER_MAX_TOKENS, REASONING_SAFE_MAX_TOKENS) : ANSWER_MAX_TOKENS;
+
+	const response = await completeSimple(
+		model,
+		{
+			systemPrompt: [DIFFICULTY_SYSTEM_PROMPT],
+			messages: [{ role: "user", content: input, timestamp: Date.now() }],
+		},
+		{
+			apiKey,
+			maxTokens,
+			disableReasoning: true,
+			metadata,
+			signal: deps.signal,
+		},
+	);
+
+	if (response.stopReason === "error") {
+		throw new Error(`auto-thinking: online classification failed: ${response.errorMessage ?? "unknown error"}`);
+	}
+
+	const text = extractText(response.content);
+	const effort = parseDifficultyLevel(text);
+	if (!effort) {
+		throw new Error(`auto-thinking: unparseable online classification: ${JSON.stringify(text)}`);
+	}
+	return effort;
+}
+
+async function classifyLocal(input: string, modelKey: string, deps: ClassifyDifficultyDeps): Promise<Effort> {
+	if (!isTinyMemoryLocalModelKey(modelKey)) {
+		throw new Error(`auto-thinking: unsupported local classifier model: ${modelKey}`);
+	}
+	const builtPrompt = prompt.render(difficultyLocalPrompt, { prompt: input });
+	const text = await tinyModelClient.complete(modelKey, builtPrompt, {
+		maxTokens: ANSWER_MAX_TOKENS,
+		signal: deps.signal,
+	});
+	if (!text) {
+		throw new Error("auto-thinking: local classification returned no output");
+	}
+	const effort = parseDifficultyBucket(text);
+	if (!effort) {
+		throw new Error(`auto-thinking: unparseable local classification: ${JSON.stringify(text)}`);
+	}
+	return effort;
+}
+
+/** Map the online 4-way level keyword to an {@link Effort}; earliest match wins. */
+export function parseDifficultyLevel(text: string): Effort | undefined {
+	const lower = text.toLowerCase();
+	const candidates: Array<[number, Effort]> = [];
+	// `xhigh` must be probed as its own token: `\bhigh\b` cannot match the "high"
+	// inside "xhigh" (no word boundary between `x` and `h`), so the two never collide.
+	const xhigh = lower.search(/x[\s_-]?high/);
+	if (xhigh >= 0) candidates.push([xhigh, Effort.XHigh]);
+	const high = lower.search(/\bhigh\b/);
+	if (high >= 0) candidates.push([high, Effort.High]);
+	const medium = lower.search(/\bmed(?:ium)?\b/);
+	if (medium >= 0) candidates.push([medium, Effort.Medium]);
+	const low = lower.search(/\blow\b/);
+	if (low >= 0) candidates.push([low, Effort.Low]);
+	return earliest(candidates);
+}
+
+/** Map the local 3-way bucket keyword to an {@link Effort}; earliest match wins. */
+export function parseDifficultyBucket(text: string): Effort | undefined {
+	const lower = text.toLowerCase();
+	const candidates: Array<[number, Effort]> = [];
+	const trivial = lower.search(/\btrivial\b/);
+	if (trivial >= 0) candidates.push([trivial, Effort.Low]);
+	const moderate = lower.search(/\bmoderate\b/);
+	if (moderate >= 0) candidates.push([moderate, Effort.High]);
+	const hard = lower.search(/\bhard\b/);
+	if (hard >= 0) candidates.push([hard, Effort.XHigh]);
+	return earliest(candidates);
+}
+
+function earliest(candidates: Array<[number, Effort]>): Effort | undefined {
+	if (candidates.length === 0) return undefined;
+	let best = candidates[0];
+	for (const candidate of candidates) {
+		if (candidate[0] < best[0]) best = candidate;
+	}
+	return best[1];
+}
+
+function extractText(content: AssistantMessage["content"]): string {
+	return content
+		.filter((block): block is Extract<AssistantMessage["content"][number], { type: "text" }> => block.type === "text")
+		.map(block => block.text)
+		.join(" ")
+		.trim();
+}
+
+/**
+ * Bound the classifier input. Code blocks are kept (a large diff is signal), but
+ * very long prompts are head+tail trimmed so the intent (start) and any trailing
+ * error/stacktrace (end) both survive.
+ */
+function prepareClassifierInput(text: string): string {
+	const trimmed = text.trim();
+	if (trimmed.length <= MAX_INPUT_CHARS) return trimmed;
+	return `${trimmed.slice(0, HEAD_CHARS)}\n…\n${trimmed.slice(-TAIL_CHARS)}`;
+}